Symmetric autoencoder using PyTorch with document classifier and data-scrapper for Wikipedia articles.
Preliminary results are curated in the Jupiter notebook, Results.ipynb, using two sets of data (a pilot of 10 classes * 100 articles per class, and full data 10 classes * 1000 articles) and two preprocessing methods (word frequencies, word embeddings). A couple of baselines are also included for comparison.
There is still a lot of room for hyper-parameter optimization. But the essentials are done. The model architecture is expandable by further work, through more complex architectures.
The size of the model is very manageable, but the reconstructed feature set is big. Data for most stages are included in /data directory: from raw scrapped text data (using the scrapper class in /code/) to processed data in numeric format with two mechanisms: either word embeddings or frequencies.
You can run the code to train the model and do a basic classification. Configuration can be changed in the run_model.py script.
python run_model.py
Mohammad Yaghini, April 2018